X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=ffmpeg_capture.cpp;h=7bd927801359b5906db8cec4641557c1898a82ed;hb=6ffaabac0a523617b686f40c154a25cb548cc561;hp=9a151b60d044c9eb038cdcc7a8c74b5a98e7bbd5;hpb=804bfa72263942d9beebe749ef0c8dc07a811413;p=nageru diff --git a/ffmpeg_capture.cpp b/ffmpeg_capture.cpp index 9a151b6..7bd9278 100644 --- a/ffmpeg_capture.cpp +++ b/ffmpeg_capture.cpp @@ -208,9 +208,10 @@ YCbCrFormat decode_ycbcr_format(const AVPixFmtDescriptor *desc, const AVFrame *f FFmpegCapture::FFmpegCapture(const string &filename, unsigned width, unsigned height) : filename(filename), width(width), height(height), video_timebase{1, 1} { - // Not really used for anything. description = "Video: " + filename; + last_frame = steady_clock::now(); + avformat_network_init(); // In case someone wants this. } @@ -281,13 +282,20 @@ void FFmpegCapture::producer_thread_func() pthread_setname_np(pthread_self(), thread_name); while (!producer_thread_should_quit.should_quit()) { - string pathname = search_for_file(filename); - if (filename.empty()) { - fprintf(stderr, "%s not found, sleeping one second and trying again...\n", filename.c_str()); + string filename_copy; + { + lock_guard lock(filename_mu); + filename_copy = filename; + } + + string pathname = search_for_file(filename_copy); + if (pathname.empty()) { + fprintf(stderr, "%s not found, sleeping one second and trying again...\n", filename_copy.c_str()); send_disconnected_frame(); producer_thread_should_quit.sleep_for(seconds(1)); continue; } + should_interrupt = false; if (!play_video(pathname)) { // Error. fprintf(stderr, "Error when playing %s, sleeping one second and trying again...\n", pathname.c_str()); @@ -313,17 +321,33 @@ void FFmpegCapture::send_disconnected_frame() VideoFormat video_format; video_format.width = width; video_format.height = height; - video_format.stride = width * 4; video_format.frame_rate_nom = 60; video_format.frame_rate_den = 1; video_format.is_connected = false; - - video_frame.len = width * height * 4; - memset(video_frame.data, 0, video_frame.len); + if (pixel_format == bmusb::PixelFormat_8BitBGRA) { + video_format.stride = width * 4; + video_frame.len = width * height * 4; + memset(video_frame.data, 0, video_frame.len); + } else { + video_format.stride = width; + current_frame_ycbcr_format.luma_coefficients = YCBCR_REC_709; + current_frame_ycbcr_format.full_range = true; + current_frame_ycbcr_format.num_levels = 256; + current_frame_ycbcr_format.chroma_subsampling_x = 2; + current_frame_ycbcr_format.chroma_subsampling_y = 2; + current_frame_ycbcr_format.cb_x_position = 0.0f; + current_frame_ycbcr_format.cb_y_position = 0.0f; + current_frame_ycbcr_format.cr_x_position = 0.0f; + current_frame_ycbcr_format.cr_y_position = 0.0f; + video_frame.len = width * height * 2; + memset(video_frame.data, 0, width * height); + memset(video_frame.data + width * height, 128, width * height); // Valid for both NV12 and planar. + } frame_callback(-1, AVRational{1, TIMEBASE}, -1, AVRational{1, TIMEBASE}, timecode++, video_frame, /*video_offset=*/0, video_format, FrameAllocator::Frame(), /*audio_offset=*/0, AudioFormat()); + last_frame_was_connected = false; } } @@ -341,7 +365,10 @@ bool FFmpegCapture::play_video(const string &pathname) last_modified = buf.st_mtim; } - auto format_ctx = avformat_open_input_unique(pathname.c_str(), nullptr, nullptr); + AVDictionary *opts = nullptr; + av_dict_set(&opts, "fflags", "nobuffer", 0); + + auto format_ctx = avformat_open_input_unique(pathname.c_str(), nullptr, &opts, AVIOInterruptCB{ &FFmpegCapture::interrupt_cb_thunk, this }); if (format_ctx == nullptr) { fprintf(stderr, "%s: Error opening file\n", pathname.c_str()); return false; @@ -381,8 +408,9 @@ bool FFmpegCapture::play_video(const string &pathname) video_codec_ctx.get(), avcodec_close); // Open audio decoder, if we have audio. - AVCodecContextWithDeleter audio_codec_ctx = avcodec_alloc_context3_unique(nullptr); + AVCodecContextWithDeleter audio_codec_ctx; if (audio_stream_index != -1) { + audio_codec_ctx = avcodec_alloc_context3_unique(nullptr); const AVCodecParameters *audio_codecpar = format_ctx->streams[audio_stream_index]->codecpar; audio_timebase = format_ctx->streams[audio_stream_index]->time_base; if (avcodec_parameters_to_context(audio_codec_ctx.get(), audio_codecpar) < 0) { @@ -405,6 +433,7 @@ bool FFmpegCapture::play_video(const string &pathname) internal_rewind(); // Main loop. + bool first_frame = true; while (!producer_thread_should_quit.should_quit()) { if (process_queued_commands(format_ctx.get(), pathname, last_modified, /*rewound=*/nullptr)) { return true; @@ -453,15 +482,72 @@ bool FFmpegCapture::play_video(const string &pathname) pts_origin = frame->pts; } next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate); + if (first_frame && last_frame_was_connected) { + // If reconnect took more than one second, this is probably a live feed, + // and we should reset the resampler. (Or the rate is really, really low, + // in which case a reset on the first frame is fine anyway.) + if (duration(next_frame_start - last_frame).count() >= 1.0) { + last_frame_was_connected = false; + } + } video_frame->received_timestamp = next_frame_start; + + // The easiest way to get all the rate conversions etc. right is to move the + // audio PTS into the video PTS timebase and go from there. (We'll get some + // rounding issues, but they should not be a big problem.) + int64_t audio_pts_as_video_pts = av_rescale_q(audio_pts, audio_timebase, video_timebase); + audio_frame->received_timestamp = compute_frame_start(audio_pts_as_video_pts, pts_origin, video_timebase, start, rate); + + if (audio_frame->len != 0) { + // The received timestamps in Nageru are measured after we've just received the frame. + // However, pts (especially audio pts) is at the _beginning_ of the frame. + // If we have locked audio, the distinction doesn't really matter, as pts is + // on a relative scale and a fixed offset is fine. But if we don't, we will have + // a different number of samples each time, which will cause huge audio jitter + // and throw off the resampler. + // + // In a sense, we should have compensated by adding the frame and audio lengths + // to video_frame->received_timestamp and audio_frame->received_timestamp respectively, + // but that would mean extra waiting in sleep_until(). All we need is that they + // are correct relative to each other, though (and to the other frames we send), + // so just align the end of the audio frame, and we're fine. + size_t num_samples = (audio_frame->len * 8) / audio_format.bits_per_sample / audio_format.num_channels; + double offset = double(num_samples) / OUTPUT_FREQUENCY - + double(video_format.frame_rate_den) / video_format.frame_rate_nom; + audio_frame->received_timestamp += duration_cast(duration(offset)); + } + + steady_clock::time_point now = steady_clock::now(); + if (duration(now - next_frame_start).count() >= 0.1) { + // If we don't have enough CPU to keep up, or if we have a live stream + // where the initial origin was somehow wrong, we could be behind indefinitely. + // In particular, this will give the audio resampler problems as it tries + // to speed up to reduce the delay, hitting the low end of the buffer every time. + fprintf(stderr, "%s: Playback %.0f ms behind, resetting time scale\n", + pathname.c_str(), + 1e3 * duration(now - next_frame_start).count()); + pts_origin = frame->pts; + start = next_frame_start = now; + timecode += MAX_FPS * 2 + 1; + } bool finished_wakeup = producer_thread_should_quit.sleep_until(next_frame_start); if (finished_wakeup) { if (audio_frame->len > 0) { assert(audio_pts != -1); } + if (!last_frame_was_connected) { + // We're recovering from an error (or really slow load, see above). + // Make sure to get the audio resampler reset. (This is a hack; + // ideally, the frame callback should just accept a way to signal + // audio discontinuity.) + timecode += MAX_FPS * 2 + 1; + } frame_callback(frame->pts, video_timebase, audio_pts, audio_timebase, timecode++, video_frame.get_and_release(), 0, video_format, audio_frame.get_and_release(), 0, audio_format); + first_frame = false; + last_frame = steady_clock::now(); + last_frame_was_connected = true; break; } else { if (producer_thread_should_quit.should_quit()) break; @@ -550,6 +636,7 @@ AVFrameWithDeleter FFmpegCapture::decode_frame(AVFormatContext *format_ctx, AVCo AVFrameWithDeleter video_avframe = av_frame_alloc_unique(); bool eof = false; *audio_pts = -1; + bool has_audio = false; do { AVPacket pkt; unique_ptr pkt_cleanup( @@ -568,9 +655,7 @@ AVFrameWithDeleter FFmpegCapture::decode_frame(AVFormatContext *format_ctx, AVCo return AVFrameWithDeleter(nullptr); } } else if (pkt.stream_index == audio_stream_index) { - if (*audio_pts == -1) { - *audio_pts = pkt.pts; - } + has_audio = true; if (avcodec_send_packet(audio_codec_ctx, &pkt) < 0) { fprintf(stderr, "%s: Cannot send packet to audio codec.\n", pathname.c_str()); *error = true; @@ -582,10 +667,13 @@ AVFrameWithDeleter FFmpegCapture::decode_frame(AVFormatContext *format_ctx, AVCo } // Decode audio, if any. - if (*audio_pts != -1) { + if (has_audio) { for ( ;; ) { int err = avcodec_receive_frame(audio_codec_ctx, audio_avframe.get()); if (err == 0) { + if (*audio_pts == -1) { + *audio_pts = audio_avframe->pts; + } convert_audio(audio_avframe.get(), audio_frame, audio_format); } else if (err == AVERROR(EAGAIN)) { break; @@ -645,10 +733,15 @@ void FFmpegCapture::convert_audio(const AVFrame *audio_avframe, FrameAllocator:: } audio_format->num_channels = 2; + int64_t channel_layout = audio_avframe->channel_layout; + if (channel_layout == 0) { + channel_layout = av_get_default_channel_layout(audio_avframe->channels); + } + if (resampler == nullptr || audio_avframe->format != last_src_format || dst_format != last_dst_format || - av_frame_get_channel_layout(audio_avframe) != last_channel_layout || + channel_layout != last_channel_layout || av_frame_get_sample_rate(audio_avframe) != last_sample_rate) { avresample_free(&resampler); resampler = avresample_alloc_context(); @@ -657,8 +750,8 @@ void FFmpegCapture::convert_audio(const AVFrame *audio_avframe, FrameAllocator:: exit(1); } - av_opt_set_int(resampler, "in_channel_layout", av_frame_get_channel_layout(audio_avframe), 0); - av_opt_set_int(resampler, "out_channel_layout", AV_CH_LAYOUT_STEREO, 0); + av_opt_set_int(resampler, "in_channel_layout", channel_layout, 0); + av_opt_set_int(resampler, "out_channel_layout", AV_CH_LAYOUT_STEREO_DOWNMIX, 0); av_opt_set_int(resampler, "in_sample_rate", av_frame_get_sample_rate(audio_avframe), 0); av_opt_set_int(resampler, "out_sample_rate", OUTPUT_FREQUENCY, 0); av_opt_set_int(resampler, "in_sample_fmt", audio_avframe->format, 0); @@ -671,7 +764,7 @@ void FFmpegCapture::convert_audio(const AVFrame *audio_avframe, FrameAllocator:: last_src_format = AVSampleFormat(audio_avframe->format); last_dst_format = dst_format; - last_channel_layout = av_frame_get_channel_layout(audio_avframe); + last_channel_layout = channel_layout; last_sample_rate = av_frame_get_sample_rate(audio_avframe); } @@ -680,7 +773,7 @@ void FFmpegCapture::convert_audio(const AVFrame *audio_avframe, FrameAllocator:: uint8_t *data = audio_frame->data + audio_frame->len; int out_samples = avresample_convert(resampler, &data, 0, num_samples_room, - audio_avframe->data, audio_avframe->linesize[0], audio_avframe->nb_samples); + const_cast(audio_avframe->data), audio_avframe->linesize[0], audio_avframe->nb_samples); if (out_samples < 0) { fprintf(stderr, "Audio conversion failed.\n"); exit(1); @@ -783,3 +876,13 @@ UniqueFrame FFmpegCapture::make_video_frame(const AVFrame *frame, const string & return video_frame; } + +int FFmpegCapture::interrupt_cb_thunk(void *unique) +{ + return reinterpret_cast(unique)->interrupt_cb(); +} + +int FFmpegCapture::interrupt_cb() +{ + return should_interrupt.load(); +}